Autonomous Skill Contract

Added: 2026-05-08 Spec: .correctless/specs/autonomous-skill-contract.md Intensity: high

What It Does

Every skill now declares how it behaves when dispatched by /cauto: interactively (pausing for human input), autonomously (applying sensible defaults), or hybrid (some decisions autonomous, some always-escalate). The interaction_mode frontmatter field in each SKILL.md encodes this contract.

When /cauto runs a skill, it passes mode: autonomous in the task prompt. Skills detect this flag and use their documented defaults instead of pausing. Every autonomous decision is returned as structured output and logged to a JSONL artifact for end-of-pipeline review.

How It Works 

graph LR
    A["/cauto dispatches skill"] --> B{"mode: autonomous<br/>in prompt?"}
    B -->|yes| C["Use Autonomous Defaults<br/>(AD-001, AD-002, ...)"]
    B -->|no| D["Interactive mode<br/>pause for human input"]
    C --> E["Return decisions as<br/>structured output"]
    E --> F["/cauto writes to<br/>decisions JSONL (ABS-030)"]
    F --> G["End-of-pipeline summary"]
    G --> H{"Deferred escalations?"}
    H -->|yes| I["Human confirmation<br/>before PR creation"]
    H -->|no| J["PR creation proceeds"]

    style C fill:#51cf66,color:#fff
    style D fill:#339af0,color:#fff
    style I fill:#ffd43b,color:#000

Skill Classification

Mode Count Description Examples
autonomous 5 Run to completion without human input chelp, cmetrics, cstatus, csummary, cwtf
interactive 2 Require Socratic human interaction csetup, cspec
hybrid 22 Have decision points with defaults; some always-escalate ctdd, cdocs, cverify, caudit, cauto, …

Four hybrid skills also have context: fork (cdevadv, cpostmortem, credteam, cverify). These cannot receive follow-up input when forked, so escalate: always decisions use the deferred escalation mechanism (R-011): apply the default, flag it, and surface it at pipeline conclusion for human review.

Decision Logging

/cauto is the sole writer of .correctless/artifacts/autonomous-decisions-{branch_slug}.jsonl (ABS-030). Skills return decisions in a structured block; /cauto persists them via scripts/autonomous-decision-writer.sh. The writer script follows the same SFG-bypass pattern as audit-record.sh (ABS-029).

Each entry records: skill name, decision ID (AD-xxx or AD-UNLISTED-N), default applied, rationale, whether escalation was deferred, and the original escalation reason if applicable.

Configuration

No configuration required. The interaction_mode field is documentation-only – it is not parsed by the Claude Code plugin loader (ENV-007). /cauto reads it via the Read tool when determining dispatch behavior.

Known Limitations

  • AD-UNLISTED accumulation: If many decisions hit the AD-UNLISTED fallback (R-014), the skill’s ## Autonomous Defaults section is incomplete. Unlisted decisions appear prominently in the deferred escalation summary.
  • Default quality: Autonomous defaults may become stale as the project evolves. Defaults are visible in interactive mode too (R-008 (default) annotations), providing natural feedback.

References

  • Spec: .correctless/specs/autonomous-skill-contract.md (14 rules)
  • ABS-030 in .correctless/ARCHITECTURE.md
  • Tests: tests/test-autonomous-skill-contract.sh (70 assertions)
  • Writer script: scripts/autonomous-decision-writer.sh